content-repurposer

# Content Repurposer One video, edited once, becomes three platform-correct posts — and TikTok + YouTube get **scheduled automatically**. The thing that breaks when you cross-post is the **call-to-action**: "comment WORD and I'll send the link" is right on Instagram (DMs are the mechanic) but means nothing on TikTok or YouTube, where the link lives in the bio / description. This skill detects the CTA and **translates** it per platform — on the video and in the caption — then schedules the result. The source file is **read-only**. Every output is a new file in a separate folder. --- ## Phase 0 — Setup (first run only) Before anything else, check whether setup has been done. Look for **both** `repurpose-config.md` and `.env` at the skill root. - If either is missing → run setup now (below). - If both exist → skip straight to Phase 1. **Setup steps:** 1. **Install the tools.** Run: ```bash bash scripts/setup.sh ``` This installs `ffmpeg`, `tesseract`, `faster-whisper`, and `Pillow` (all free, all local — `faster-whisper` is the no-key, no-cost transcription engine) and creates `.env` from the template. If it reports a failure, walk the user through `references/setup-guide.md` for their OS, then re-run it. 2. **Connect Zernio.** Read `references/zernio-setup.md` to the user and have them: create a Zernio account → connect **TikTok + YouTube** (first 2 accounts free) → copy their API key (`sk_…`) from Settings → API Keys → paste it into `.env` as `ZERNIO_API_KEY=…`. Confirm with: ```bash python scripts/schedule_zernio.py --list-accounts ``` 3. **Write their config.** Interview the user for the values in `assets/repurpose-config.example.md` (handle, the link, default trigger word, timezone + default post time, TikTok privacy, YouTube visibility, whether Instagram is connected). Copy the example to `repurpose-config.md` at the skill root and fill it in. Paste the TikTok/YouTube account IDs from the `--list-accounts` output so later runs skip the lookup. Show a short summary to confirm. Re-customizing later = editing `repurpose-config.md` or `.env`. Never edit code. --- ## Phase 1 — Analyze the video Run **both** passes on the source (audio CTAs and on-screen-text CTAs are different problems — always do both): ```bash # Spoken CTA: transcript + classification + trigger word, into a report python scripts/repurpose.py "PATH/TO/video.mp4" --out-dir output --dry-run # On-screen text CTA: OCR the final seconds python scripts/scan_text_cta.py "PATH/TO/video.mp4" ``` From the two reports, decide **where the CTA lives** — this drives everything: - **Spoken** (in the voiceover) → `repurpose.py` found it (`cta_type`, `trigger_word`). - **On-screen text** (a card / overlay) → `scan_text_cta.py` found it (`card_start`, `type`, `trigger`). - **Description/caption only** (nothing on the video) → both report nothing; the video is reused as-is and only the captions change. Also note the **source platform** (ask if unclear). The usual case is an Instagram-style master (ends with "comment WORD"). If the source is already a TikTok-style "link in bio" video, the translation runs the other direction (see the table below) — the mechanics are symmetric. --- ## Phase 2 — Translate the CTA per platform (core logic) The link delivery differs per platform, so the CTA must change to match: | Platform | CTA the video/caption should end on | |----------|--------------------------------------| | **Instagram** | "comment {TRIGGER} and I'll send you the link" (the DM mechanic) | | **TikTok** | "Link in bio" | | **YouTube** | "Link in description" | How you produce each platform's **video** depends on where the CTA lived: ### A. Spoken CTA Use `repurpose.py`'s trim. Instagram **keeps** the spoken comment-CTA; TikTok and YouTube get the tail trimmed (the CTA then lives only in their captions). The defaults already do this: ```bash python scripts/repurpose.py "PATH/TO/video.mp4" --out-dir output # -> video__tiktok.mp4, video__youtube_shorts.mp4 (CTA trimmed), + report.json # Instagram version = the original master (CTA kept). ``` Confirm the cut in `report.json` (`cut_if_removed` lands after the last real sentence) before trusting the encode. ### B. On-screen text CTA First look at the actual frame (`ffmpeg -ss <card_start> -i video.mp4 -frames:v 1 frame.png`) and get the caption's pixel box via Tesseract TSV (see the recipe in the original engine docs / `scan_text_cta.py`). Then: - **Standalone trailing card** (CTA is the whole screen) → treat like a spoken CTA: trim for TikTok/YouTube with `--cta-at <card_start>`; keep it for Instagram. - **Caption overlaid on real content** (the video keeps playing under the card) → do **not** trim. Cover and replace the words per platform with `cover_cta.py`, one render per platform because the text differs: ```bash # TikTok python scripts/cover_cta.py video.mp4 output/video__tiktok.mp4 \ --box X Y W H --window T0 T1 --line1 "Link in bio" --line2 "for the full guide" # YouTube python scripts/cover_cta.py video.mp4 output/video__youtube.mp4 \ --box X Y W H --window T0 T1 --line1 "Link in description" --line2 "for the full guide" # Instagram keeps the original "comment {TRIGGER}" card (use the master as-is), # OR re-render it to the exact config phrasing if needed. ``` Always extract a frame from each output and view it to confirm the old text is fully covered. ### C. Description/caption-only CTA No video edit. The same master video file is scheduled to all platforms; only the caption text changes (Phase 3). Name outputs `*__instagram.mp4`, `*__tiktok.mp4`, `*__youtube.mp4`. If a platform needs no edit (e.g. Instagram in cases A/C), reuse the original master for it. --- ## Phase 3 — Write the captions + descriptions Using `references/caption-patterns.md`, `report.json` (real offer + trigger word), and the voice sample in `repurpose-config.md`, write one caption per platform with the right CTA baked in: - **Instagram** — "comment {TRIGGER} and I'll send you the link". - **TikTok** — "link in bio 🔗". - **YouTube** — a **title** (≤100 chars) **and** a description with the link. Keep the hook and topic identical across all three; only the CTA line changes. Save them to `output/<name>__captions.md` so they're easy to copy. --- ## Phase 4 — Schedule TikTok + YouTube (Zernio) Build a `job.json` (see `assets/job.example.json`) listing the TikTok and YouTube posts — each with its `file`, `content` (caption/description), `scheduledFor` (ISO 8601, derived from the user's requested date or the config's default time), plus the YouTube `title`/`visibility` and TikTok privacy from `repurpose-config.md`. Pull `timezone` from the config. Validate first, then schedule: ```bash python scripts/schedule_zernio.py --job job.json --dry-run # check payloads python scripts/schedule_zernio.py --job job.json # upload + schedule ``` The script uploads each video (presigned URL), creates one scheduled post per platform, and prints a table (platform · when · status · post id). **Instagram** is **manual by default**: print the Instagram caption + the video path and tell the user to post it themselves. Only schedule Instagram if the config says it's connected (Zernio premium) — then add an `instagram` post to `job.json`. --- ## Reporting back Give the user, per platform: where the CTA ended up (comment / bio / description), the output path, and the scheduled time + Zernio post id (or "manual" for Instagram). Then paste the three captions. Keep prose short — they can watch the files themselves. ## Notes and limits - CTA detection is keyword-based (transcript + OCR), tuned for "comment X", "link in bio", "follow me" endings. Novel phrasing may need `repurpose.py --cta-at`. Patterns live at the top of `repurpose.py` / `scan_text_cta.py`. - Run **both** detection passes — audio detection misses silent on-screen text CTAs. - `cover_cta.py` cleanly replaces a solid caption box; it's not built for text baked into busy footage (moving text, text over a face) — surface a frame and ask. - Zernio's first 2 connected accounts are free (TikTok + YouTube). Instagram auto-scheduling needs a 3rd (paid) account; otherwise it stays manual. - TikTok may restrict which `privacy_level` the API can use until the account is approved for public posting — if a post is rejected, lower it in the config.